This work uses data-driven approaches to study the feasibility of reconstructing ocean interior variables (temperature and salinity) from surface observations provided by satellites and interior observations provided by buoys. The feasibility of the approach is based on an Observing System Simulation Experiment (OSSE) in which we use the outputs from an ocean numerical model as the ground truth, and simulate a real observing system of the ocean, taking the surface of the model as a simulation of satellite observations, and vertical profiles in the same locations as the real buoys. We implemented different models based on Random Forest Regressors and Long-Short Term Memory networks which were trained with the simulated observations and validated against the complete numerical model results. We obtain high spatial and temporal correlation using both technologies and an accurate description of the annual variability of the data accompanied by small biases.